Making System Calls Transaction Safe

The advent of multi-core processors has renewed interest in the idea of incorporating transactions into the programming model used to write multi-threaded programs. However, using lock based multi-threaded programming is hard to get right and easy to deadlock. Transactional Memory (TM) simplifies multi-threaded programming, by replacing use of locks with transactions. But, there are many problems that TM system development is facing today. Firstly, there are limited TM workloads, since TM systems existing today are in a preliminary stage. Secondly, TM systems cannot handle system calls inside transactions as of today. In this project, we look at TM related issues on two important server workloads AOLserver and BIND. Our focus is primarily on network send()/recv() calls and their variants. We propose using a cross-layer mechanism which leverages both kernel and user level support to provide an easy and minimum performance overhead solution to handle network calls. From our current results we see upto 3X overhead when send()/recv() are used for sending/receiving large packets (e.g. 1400 bytes). But, we believe that can be brought down to a negligible level using simple optimizations. Finally, we also handle conditional sychronization in transactions which can potentially lead to a deadlock. We use a simple busy-wait approach explored later in the paper.